- The discussion revolves around what makes someone human, with the answers being humor and emotion.
- The conversation then shifts to the advancement of AI and AGI (Artificial General Intelligence).
- The speakers discuss the evolution of AI and AGI, with the definition of intelligence constantly changing.
- They predict that AGI will be achieved within the next decade, but the specific definition of AGI is still being debated.
- The goal is to create AGI that surpasses human capabilities in various domains.
- The speakers emphasize the importance of abundant and inexpensive intelligence in improving the human condition.
- The conversation also touches on the challenges and benefits of AGI, as well as the need for reliability and safety in AI systems.

- OpenAI is focused on increasing the reliability and safety of their AI systems while also improving their capabilities.
- They are working on reducing the issue of hallucination in their models, and GPT-4 has made progress in this area.
- OpenAI is exploring different partnerships and data sources to improve the quality of training data for their models.
- They believe that as the models become smarter, they will require less training data.
- The core value of GPT models is their reasoning capacity, not just their ability to memorize data.
- OpenAI has a good relationship with Microsoft, although there may be some competition in certain areas.
- OpenAI is introducing more personalization to their models, allowing for custom instructions and voice recognition.
- The goal is to build deep relationships with these AI bots that enhance productivity and creativity while being reliable and personalized.

- The goal is for AI to become more personalized, useful, fun, and creative
- There will be multiple AI systems personalized for specific domains and tasks
- The decisions about AI control belong to society as a whole
- The development of AI will go through societal discussions, competition, regulatory measures, and other forces
- The speaker has misgivings about a future where AI friends replace human friends, but acknowledges that others may want it
- Personalization is important, but it should be clear when interacting with AI and when not
- The ideal device for AI interaction is still being explored
- Custom hardware may not be necessary, partnerships with other companies are in place
- The company will need more capital to complete its mission
- The product side of the business is growing the most
- Revenue comes from both the platform side and first-party collaborations
- The company wants to offer more affordable options, such as family plans, in the future
- There are concerns about the potential misuse of AI, but the speaker believes in addressing the risks while pursuing the positive impact
- It is important to be thoughtful about the risks and mitigate them in AI technology development.

- The speakers emphasize the importance of being thoughtful about the risks and building technology that mitigates those risks.
- They highlight the need to understand and address safety risks in various aspects of AI, including data, models, tools, and policies.
- The speakers believe that deploying AI in the real world and adapting to its implications is crucial for safe deployment.
- They mention the importance of self-regulation and the role of governments in coordinating international regulation for advanced AI models.
- The responsibility for detecting and addressing AI-generated content lies with the companies developing the technology, but it also requires input from users and society.
- OpenAI is working on technologies to detect AI-generated images and texts, but they are still testing and refining them.
- The release of these technologies will be done in a controlled manner, with feedback and improvements along the way.
- The watermarking of content is a nuanced issue, and OpenAI is willing to follow the collective wishes of society in this regard.
- The speakers highlight the potential risks of individualized persuasion as AI advances, which poses challenges in dealing with the impact on society.

- The speaker believes that the future will be great and that AI is one of the most significant inventions in human history.
- They acknowledge that there are fears and uncertainties about the impact of AI, but they are hopeful about the positive outcomes.
- The speaker dismisses the question about having a bunker as a ridiculous one.
- They discuss the hope of pushing civilization forward through augmented collective intelligence.
- The speaker acknowledges that the nature of work will change due to AI, and it is important to involve society in the discussion and preparation for these changes.
- They emphasize the need for accessibility and interaction with AI technology to understand its potential impact.
- The speaker believes that although jobs may change, society will adapt and find new and better jobs.
- They stress the importance of addressing the speed at which job market changes will occur and ensuring that people have agency and the ability to influence the future.
- The speaker highlights the satisfaction and innate desire for creative expression and contribution to society.
- They state the need for joint collaboration in shaping the future and making AI technology widely available for people to participate and push the future forward.

- The discussion explores what makes someone human, focusing on humor and emotion.
- The advancement of AI and AGI is discussed, with predictions that AGI will be achieved in the next decade.
- OpenAI is working on improving the reliability and safety of their AI systems while also enhancing their capabilities.
- The company is exploring partnerships and data sources to improve training data for their models.
- OpenAI aims to personalize their AI models for enhanced productivity and creativity.
- AI control decisions belong to society as a whole, and the development of AI will involve societal discussions, competition, and regulation.
- The potential risks of AI misuse are acknowledged, and OpenAI is committed to addressing them while pursuing positive impact.
- The responsibility for detecting and addressing AI-generated content lies with the companies, with input from users and society.
- OpenAI is working on technologies to detect AI-generated images and texts and will release them in a controlled manner.
- The speaker believes in the positive outcomes of AI and emphasizes the need for societal involvement in preparing for job market changes.
- The speaker highlights the importance of accessibility, collaboration, and creative expression in shaping the future with AI.

OpenAI's Sam Altman and Mira Murati discuss the capabilities of their future GPT models, how human relationships with AI will change in the future and fears about safety, liability and work as the technology advances. 

#OpenAI #ChatGPT #WSJ

You both have to answer what makes you human.

To confirm you're both human, I'm going to need you to confirm which of these boxes have a traffic light.

Well, Sam, you were actually here 9 years ago at our first Tech Live.

And I actually want to roll the clip of what you said.

Certainly the fear with AI or machine intelligence in general is that it replaces drivers or doctors or whatever.

But the optimistic view on this, and certainly what backs up what we're seeing, is that computers and humans are very good at very different things.

So a computer doctor will outcrunch the numbers and do a better job than a human on looking at a massive amount of data and saying this.

But on cases that require judgment or creativity or empathy, we are nowhere near any computer system that is any good at this.

People I think the prevailing wisdom back then was that AI was going to do the kind of like robotic jobs really well first.

So it would have been a great robotic surgeon, something like that.

And then maybe eventually it was gonna do the sort of like higher judgment tasks.

And then, you know, then it would kind of do the empathy and then maybe, never, it was gonna be like a really great creative thinker.

And creativity has been in some sense, and at this point the definition of the word creativity is up for debate, but creativity in some sense has been easier for AI than people thought.

You can see Dolly 3 generate these amazing images or write these creative stories with GPT-4 or whatever.

So that part of the answer maybe was not perfect.

And I certainly would not have predicted GPT-4 9 years ago, quite how it turned out.

But a lot of the other parts about people still really want a human doctor, that's definitely very true.

You could just define it for everybody in the audience?

I will say it's a system that can generalize across many domains that, you know, would be equivalent to human work.

They produce a lot of productivity and economic value.

But we're talking about 1 system that can generalize across a lot of digital domains of human work.

The 2 things that I think will matter most over the next decade or few decades to improving the human condition the most, giving us sort of just more of what we want, our abundant and inexpensive intelligence.

The more powerful, the more general, the smarter, the better.

I think that is AGI, and then abundant and cheap energy.

And if we can get these 2 things done in the world, then it's almost like difficult to imagine how much else we could do.

We're big believers that you give people better tools and they do things that astonish you, and I think AGI will be the best tool humanity has yet created.

With it, we will be able to solve all sorts of problems.

We'll be able to express ourselves in new creative ways.

We'll make just incredible things for each other, for ourselves, for the world, for this unfolding human story.

And it's new, and anything new comes with change, and change is not always all easy.

But I think this will be just absolutely tremendous upside.

And we're gonna, 9 more years, if you're nice enough to invite me back, you'll roll this question and people will say like, how could we have thought we didn't want this?

I guess 2 parts to that, my next question, when will it be here and how will we know it's here?

I mean, you can both predict how long, and we'll call you in 10 years

I mean, yeah, probably in the next decade.

But I would say It's a bit tricky because we, you know, when will it be here, right?

And I just kind of gave you a definition, but then often we talk about intelligence and, you know, how intelligent is it or whether it's conscious and sentient and all of these terms.

And, you know, they're not quite right because they sort of define our own intelligence, and we're building something slightly different.

And you can kind of see how the definition of intelligence evolves from, you know, machines that were really great at chess and AlphaGo and now the GPT series, and then what's next, but it continues to evolve, and it pushes how we define intelligence.

We kind of define AGI as the thing we don't have quite yet.

So we've moved, I mean, there were a lot of people who would have, 10 years ago said, all right, if you could make something like GPT-4 or GPT-5 maybe, that would have been an AGI.

And now people are like, well, you know, it's like a nice little chatbot or whatever.

I think it's great that the goalposts keep getting moved.

But I think we're getting close enough to whatever that AGI threshold is going to be that we no longer get to hand wave at it and the definition is going to matter.

Sam, you've used the word, and previously when describing AGI, the term median human.

I think there are experts in areas that are going to be better than AI systems for a long period of time.

And so you could come to some area where I'm really an expert at some task, and I'll be like, all right, you know, GPT-4 is doing a horrible job there, GPT-5, 6, whatever, doing a horrible job there.

But you can come to other tasks where I'm okay, but certainly not an expert.

I'm kind of like maybe like an average of what different people in the world could do at something.

And for that, then I might look at it and say, oh, this is actually doing pretty well.

So, what we mean by that is that the, in any given area, Expert humans may, like experts in any area can just do extraordinary things and that may take us a while to be able to do with these systems.

But for kind of the more average case performance, so me doing something that I'm not very good at anyway, Maybe our future versions can help me with that a lot.

And clearly at this, you're a very expert human.

And no GPT is taking your job anytime soon.

We're not there yet, but it's kind of need-to-know basis.

I would have just said, oh, yeah, here's what's happening.

All of these steps though with GPT, right, or GPT-3, 3.5, 4 are steps towards AGI.

With each of them, are you looking for a benchmark?

Yeah, so, you know, before we had the product, we were sort of looking at academic benchmarks and how well these models were doing on academic benchmarks.

And, you know, OpenAI is known for betting on scaling, you know, throwing a ton of compute and data on these neural networks and seeing how they get better and better at predicting the next token.

But it's not that we really care about the prediction of the next token, we care about the tasks in the real world to which this correlates to.

And so That's actually what we started seeing once we put out research in the real world.

And we build out products through the API, eventually through chat GPT as well.

And so now we actually have real world examples.

We can see how our customers do in specific domains, how it moves the needle for specific businesses.

And of course, with GPT-4, we saw that it did really well in exams like SAT and LSAT and so on.

So it kind of goes to our earlier point that we're continually evolving our definition of what it means for these models to be more capable.

But as we increase the capability vector, what we really look for is reliability and safety These are very interweaved and it's very important to make systems that of course are increasingly capable but that you can truly rely on and they are robust and that you can trust the output of the system so we're kind of pushing in both of these vectors at the same time and You know as we build the next model the next set of technologies We're both betting continuing to bet on scaling, but we're also looking at, you know, this other element of multi-modality.

Because we want these models to kind of perceive the world in a similar way to how we do.

And you know, we perceive the world not just in text, but images and sounds and so on.

So we want to have robust representations of the world in these models.

Will GPT-5 solve the hallucination problem?

We've made a ton of progress on the hallucination issue with GPT-4, but we're still quite, we're not where we need to be.

But you know, we're sort of on the right track, and it's unknown, it's research.

It could be that continuing in this path of reinforcement learning with human feedback, we can get all the way to really reliable outputs.

And we're also adding other elements like retrieval and search, so you can, you have the ability to provide more factual answers or to get more factual outputs from the model.

So there's a combination of technologies that we're putting together to kind of reduce the hallucination issue.

Sam, I'll ask you about the data, the training data.

Obviously, there's been maybe some people in this audience who may not be thrilled about some of the data that you guys have used to train some of your models.

Not too far from here in Hollywood, people have not been thrilled, publishers.

When you're considering now as you're walking through and going to work towards these next models, what are the conversations you're having around the data?

So a few thoughts in different directions here.

1, we obviously only want to use data that people are excited about us using.

We want the model of this new world to work for everyone.

And we want to find ways to make people say, like, you know what, I see why this is great.

I see why this is going to be a new, it may be a new way that we think about some of these issues around data ownership and how economic flows work.

But we want to get to something that everybody feels really excited about.

But 1 of the challenges has been people, different kinds of data owners have very different pictures.

So we're just experimenting with a lot of things.

We're doing partnerships of different shapes.

And we think, like with any new field, we'll find something that just becomes a new standard.

Also, I think as these models get smarter and more capable, we will need less training data.

So I think there's this view right now which is that we're just going to like, models are going to have to train on every word humanity has ever produced or whatever.

And I technically speaking, I don't think that's what's gonna be the long-term path here.

Like we have existential proof with humans that That's not the only way to become intelligent.

And so I think the conversation gets a little bit led astray by this, because what really will matter in the future is particularly valuable data.

People want, people trust the Wall Street Journal and they want to see content from that.

And the Wall Street Journal wants that too.

And we find new models to make that work.

But I think the conversation about data and the shape of all of this, because of the technological progress we're making, it's about to shift.

Well, publishers like mine, who might be out there somewhere, they want money for that data.

Is the future of this entire race about who can pay the most for the best data?

No, that was sort of the point I was trying to make, I guess, inelegantly.

But the core, like the thing that people really like about a GPT model is not fundamentally that it has, that it knows particular knowledge.

It's that it has this larval reasoning capacity, and that's going to get better and better.

But that's really what this is going to be about.

And then there will be ways that you can set up all sorts of economic arrangements as a user or as a company making the model or whatever to say, all right, now, I understand that you would like me to go get this data from the Wall Street Journal.

I can do that, but here's the deal that's in place.

But The fundamental thing about these models is not that they memorize a lot of data.

So sort of like the model where also right now you've got Bing integrated, it goes out, looks for some of that data and can bring back some of the.

And that's, you know, on the Internet we decided, again, back in the early days of the Internet there were a lot of conversations about what The different models could be and we all kind of decided on you know here's the core framework and there's different pieces in there of course and we're all going to figure that out for AI.

Well speaking of being you in such an Adela you $10 billion friends or frenemies friends.

Yeah, yeah, I won't pretend that it's like a perfect relationship, but nowhere near the frenemy category.

You know, we get annoyed at each other about different stuff.

It just seems like increasingly, as you guys are releasing more and more products, that They seem to compete in some places.

I mean, I think that there's something core about this language interface that is a big deal.

And so there's going to be a lot of people doing things for that.

And then there's other places like we offer a version of the API.

And we work it out so that we all benefit and we're all happy and we just want like we jointly want as much usage of our models as possible.

Mira, you over the last number of months have started to roll out a lot more personalization to the model, right?

And we're going to bring down the screen to hear 1 of these clips.

But I wanted to play 1 of these conversations I had with chat GPT just yesterday.

I'm saying Chachibit is faster than this.

Yeah, yeah, but it's still that, and I'm saying that's

it's not often that people say ChachiBT is faster than X.

I've got a big interview with OpenAI executives, and I'm nervous.

Given your background as a technology columnist, you're well equipped for this.

Just be yourself, do your homework on OpenAI's latest developments, and trust in your expertise.

Remember, they'll likely appreciate thoughtful and well-researched questions.

But what's so striking about that, first of all, I want every, that's the voice of chat GPT.

It sounds so human, it sounds so natural.

It knows about me because I've already put into custom instructions, I'm a tech journalist.

Is there a future, and this is what you're maybe trying to build here, where we have deep relationships with this type of bot?

It's going to be a significant relationship, right?

Because we're building the systems that are going to be everywhere, at your home, in your educational environment, in your work environment, and maybe, you know, when you're having fun.

And so that's why it's actually so important to get it right.

And we have to be so careful about how we design this interaction so that ultimately it's, you know, elevating and it's fun and it makes productivity better and it enhances creativity.

And this is ultimately where we're trying to go.

And as we increase the capabilities of the technology, We also want to make sure that on the product side, we feel in control of these systems in the sense that we can steer them to do the things that we want them to do, and the output is reliable.

And of course, we want it to be personalized, right?

And as it has more information about your preferences, the things you like, the things you do, and the capabilities of the models increase and other features like memory and so on, it has, of course, it will become more personalized.

And that's the goal, it will become more useful.

And it's going to become more fun and more creative.

Like you can have many such systems personalized for specific domains and tasks.

And you guys will be in sort of control of people's friends, maybe people's, it gets to being people's lovers.

How do you guys think about that control?

First of all, I think we're not going to be the only player here.

We get to put our nudge on the trajectory of this technological development.

But A, we really think that the decisions belong to humanity, society as a whole, whatever you want to call it.

And B, we will be 1 of many actors building sophisticated systems here.

So it's going to be a society-wide discussion.

And there's going to be all the normal forces.

There will be competing products that offer different things.

There will be different kind of like societal embraces and pushbacks.

It's gonna be like the same complicated mess that any new technological earthing process goes through.

And then we pretty soon will turn around and we'll all feel like we had smart AI in our lives forever.

I personally have deep misgivings about this vision of the future where everyone is super close to AI friends and not more so than human friends or whatever.

I accept that other people are gonna want that.

And if that's what the world wants and what we decide makes sense, we're going to get that.

I personally think that personalization is great.

But it's important that it's not like person-ness.

And at least that you know when you're talking to AI and when you're not.

We named it ChatGBT and not, it's a long story behind that, but we named it ChatGBT and not a person's name very intentionally.

And we do a bunch of subtle things in the way you use it to, like, make it clear that you're not talking to a person.

And I think what's going to happen is that in the same way that people have a lot of relationships with people, they're going to keep doing that.

And then there will also be these, like, AIs in the world.

But you kind of know they're just a different thing.

When you're, Sam, this is another question for you, what is the ideal device that we'll interact with these on?

And I'm wondering if you, I hear you and Johnny Ive have been talking.

I think there is something great to do, but I don't know what it is yet.

I think most of the current thinking out there in the world is quite bad about what we can do with this new technology in terms of a new computing platform.

See all The Wall Street Journal transcripts on Youtube

OpenAI CEO Sam Altman and CTO Mira Murati on the Future of AI and ChatGPT | WSJ Tech Live 2023